mem-hotplug: use GFP_HIGHUSER_MOVABLE in, alloc_migrate_target()

When we offline a node, the new page should alloced from other nodes instead of the current node, because re-migrate is a waste of time.

So use prefer mempolicy for hotplug, use default mempolicy for cma.

Signed-off-by Xishi Qiu <[email protected]>

---
 include/linux/page-isolation.h | 2 +-
 mm/memory_hotplug.c            | 5 ++++-
 mm/page_alloc.c                | 2 +-
 mm/page_isolation.c            | 8 +++++---
 4 files changed, 11 insertions(+), 6 deletions(-)

include/linux/page-isolation.h +1 -1

index 047d647..c163de3 100644
@@ -65,7 +65,7 @@ undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
 			bool skip_hwpoisoned_pages);
 
-struct page *alloc_migrate_target(struct page *page, unsigned long private,
+struct page *alloc_migrate_target(struct page *page, unsigned long nid,
 				int **resultp);
 
 #endif

mm/memory_hotplug.c +3 -0

index e3cbdca..b5963bf 100644
@@ -1510,12 +1510,15 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
 	int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
 	int not_managed = 0;
 	int ret = 0;
+	int nid = NUMA_NO_NODE;
 	LIST_HEAD(source);
 
 	for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
 		if (!pfn_valid(pfn))
 			continue;
 		page = pfn_to_page(pfn);
+		if (nid == NUMA_NO_NODE)
+			nid = next_node_in(page_to_nid(page), node_online_map);
 
 		if (PageHuge(page)) {
 			struct page *head = compound_head(page);

mm/memory_hotplug.c +1 -1

 		 * migrate_pages returns # of failed pages.
 		 */
-		ret = migrate_pages(&source, alloc_migrate_target, NULL, 0,
+		ret = migrate_pages(&source, alloc_migrate_target, NULL, nid,
 					MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
 		if (ret)
 			putback_movable_pages(&source);

mm/page_alloc.c +1 -1

index 6903b69..b99f1c2 100644
@@ -7322,7 +7322,7 @@ static int __alloc_contig_migrate_range(struct compact_control *cc,
 		cc->nr_migratepages -= nr_reclaimed;
 
 		ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
-				    NULL, 0, cc->mode, MR_CMA);
+				    NULL, NUMA_NO_NODE, cc->mode, MR_CMA);
 	}
 	if (ret < 0) {
 		putback_movable_pages(&cc->migratepages);

mm/page_isolation.c +0 -0

index 4f32c9f..f471be6 100644

mm/page_isolation.c +1 -0

 	 * accordance with memory policy of the user process if possible. For
 	 * now as a simple work-around, we use the next node for destination.
+	 * Normal page: use prefer mempolicy for destination if called by

alloc_migrate_target() is called from migrate_pages(), and the page is always from user space, so we can add __GFP_HIGHMEM directly.

Quoted 2 lines · Xishi Qiu

alloc_migrate_target() is called from migrate_pages(), and the page is always from user space, so we can add __GFP_HIGHMEM directly.

No, all migratable pages are not from user space. For example, blockdev file cache has __GFP_MOVABLE and migratable but it has no __GFP_HIGHMEM and __GFP_USER.

And, zram's memory isn't GFP_HIGHUSER_MOVABLE but has __GFP_MOVABLE.

Thanks.

Quoted 2 lines · Xishi Qiu

alloc_migrate_target() is called from migrate_pages(), and the page is always from user space, so we can add __GFP_HIGHMEM directly.

No, all migratable pages are not from user space. For example, blockdev file cache has __GFP_MOVABLE and migratable but it has no __GFP_HIGHMEM and __GFP_USER.

And, zram's memory isn't GFP_HIGHUSER_MOVABLE but has __GFP_MOVABLE.

Thanks.

.

Quoted 8 lines · Joonsoo Kim

Quoted 2 lines · Xishi Qiu

alloc_migrate_target() is called from migrate_pages(), and the page is always from user space, so we can add __GFP_HIGHMEM directly.

No, all migratable pages are not from user space. For example, blockdev file cache has __GFP_MOVABLE and migratable but it has no __GFP_HIGHMEM and __GFP_USER.

Hi Joonsoo,

So the original code "gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;" is not correct?

Quoted 2 lines

And, zram's memory isn't GFP_HIGHUSER_MOVABLE but has __GFP_MOVABLE.

Can we distinguish __GFP_MOVABLE or GFP_HIGHUSER_MOVABLE when doing mem-hotplug?

Thanks,
Xishi Qiu

Quoted 5 lines

Thanks.

.

Quoted 11 lines · Xishi Qiu

Quoted 8 lines · Joonsoo Kim

Quoted 2 lines · Xishi Qiu

alloc_migrate_target() is called from migrate_pages(), and the page is always from user space, so we can add __GFP_HIGHMEM directly.

No, all migratable pages are not from user space. For example, blockdev file cache has __GFP_MOVABLE and migratable but it has no __GFP_HIGHMEM and __GFP_USER.

Quoted 4 lines

Hi Joonsoo,

So the original code "gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;" is not correct?

It's not incorrect. GFP_USER just specifies some reclaim flags, and may perhaps restrict allocation through __GFP_HARDWALL, where the original
page could have been allocated without the restriction. But it doesn't put the place in an unexpected address range, as placing a non-highmem page into highmem could. __GFP_MOVABLE then just controls a heuristic for placement within a zone.

Quoted 5 lines

Quoted 2 lines

And, zram's memory isn't GFP_HIGHUSER_MOVABLE but has __GFP_MOVABLE.

Can we distinguish __GFP_MOVABLE or GFP_HIGHUSER_MOVABLE when doing mem-hotplug?

I don't understand the question here, can you rephrase with more detail? Thanks.

Quoted 11 lines

Thanks,
Xishi Qiu

Quoted 5 lines

Thanks.

.

.

Quoted 23 lines · Vlastimil Babka

Quoted 11 lines · Xishi Qiu

Quoted 8 lines · Joonsoo Kim

Quoted 2 lines · Xishi Qiu

alloc_migrate_target() is called from migrate_pages(), and the page is always from user space, so we can add __GFP_HIGHMEM directly.

No, all migratable pages are not from user space. For example, blockdev file cache has __GFP_MOVABLE and migratable but it has no __GFP_HIGHMEM and __GFP_USER.

Quoted 4 lines

Hi Joonsoo,

So the original code "gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;" is not correct?

It's not incorrect. GFP_USER just specifies some reclaim flags, and may perhaps restrict allocation through __GFP_HARDWALL, where the original
page could have been allocated without the restriction. But it doesn't put the place in an unexpected address range, as placing a non-highmem page into highmem could. __GFP_MOVABLE then just controls a heuristic for placement within a zone.

Quoted 3 lines

Quoted 2 lines

And, zram's memory isn't GFP_HIGHUSER_MOVABLE but has __GFP_MOVABLE.

Quoted 5 lines

Quoted 2 lines

Can we distinguish __GFP_MOVABLE or GFP_HIGHUSER_MOVABLE when doing mem-hotplug?

I don't understand the question here, can you rephrase with more detail? Thanks.

Hi Joonsoo,

When we do memory offline, and the zone is movable zone,
can we use "alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);" to alloc a new page? the nid is the next node.

Quoted 25 lines · Xishi Qiu

Quoted 23 lines · Vlastimil Babka

Quoted 11 lines · Xishi Qiu

Quoted 8 lines · Joonsoo Kim

Quoted 2 lines · Xishi Qiu

alloc_migrate_target() is called from migrate_pages(), and the page is always from user space, so we can add __GFP_HIGHMEM directly.

No, all migratable pages are not from user space. For example, blockdev file cache has __GFP_MOVABLE and migratable but it has no __GFP_HIGHMEM and __GFP_USER.

Quoted 4 lines

Hi Joonsoo,

So the original code "gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;" is not correct?

It's not incorrect. GFP_USER just specifies some reclaim flags, and may perhaps restrict allocation through __GFP_HARDWALL, where the original
page could have been allocated without the restriction. But it doesn't put the place in an unexpected address range, as placing a non-highmem page into highmem could. __GFP_MOVABLE then just controls a heuristic for placement within a zone.

Quoted 3 lines

Quoted 2 lines

And, zram's memory isn't GFP_HIGHUSER_MOVABLE but has __GFP_MOVABLE.

Quoted 7 lines

Quoted 5 lines

Quoted 2 lines

Can we distinguish __GFP_MOVABLE or GFP_HIGHUSER_MOVABLE when doing mem-hotplug?

I don't understand the question here, can you rephrase with more detail? Thanks.

Hi Joonsoo,

Above is answered by Vlastimil. :)

Quoted 3 lines

When we do memory offline, and the zone is movable zone,
can we use "alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);" to alloc a new page? the nid is the next node.

I don't know much about memory offline, but, AFAIK, memory offline could happen on non-movable zone like as ZONE_NORMAL. Perhaps, you can add "if zone of the page is movable zone then alloc with GFP_HIGHUSER_MOVABLE".

Thanks.

.

Memory offline could happen on both movable zone and non-movable zone, and we can offline the whole node if the zone is movable_zone(the node only has one movable_zone), and if the zone is normal_zone, we cannot offline the whole node, because some kernel memory can't be migrated.

So how about change alloc_migrate_target() to alloc memory from the next node with GFP_HIGHUSER_MOVABLE, if the offline zone is movable_zone.